On September 22, 2026, at the United Nations, the US President announced that federal documents will refer to artificial intelligence as “Super Intelligence” — SI for short — on the argument that “artificial” understates what the technology is. No new capability shipped with the announcement. No model changed. The same week, Salesforce published the actual specification sheet for Koa, its CRM reasoning model: 120 billion parameters in a Mixture-of-Experts architecture, a one-million-token context window, and a synthetic training corpus distilled from nearly three decades of CRM deployments across more than 14 industries. These two stories look similar — both are “AI news” — and they could not be more different. One is nomenclature: it changes what documents say, and it will ripple through procurement language, vendor branding, and even domain registrations. The other is engineering: it changes what you can route to a cheaper specialized model, and when. An enterprise buyer who treats these the same will overreact to the first and underreact to the second. Here is how to respond to each.
The SI rebrand: what it touches and what it doesn’t
The renaming is administrative. It directs terminology in US government documents; it does not describe a new class of machine capability, and nothing about how models are built or governed changed with it. But terminology in federal documents is not nothing. If you sell into the public sector, expect “SI” to start appearing in RFPs and compliance language, and expect inconsistency between agencies for a while. If your contracts, security policies, or AI-use policies define “artificial intelligence” by name, a short definitions clause — “AI, also referred to as Super Intelligence or SI” — is cheap insurance against a dispute over whether a renamed term is still covered. What you should not do is rebrand your internal governance documents or your product line reactively. Renames follow hype cycles; your policies should follow capabilities.
One measurable side effect: Slovenia’s .si country domain has become the informal home of the rebrand, the way Anguilla’s .ai did a cycle earlier. Registrations jumped by thousands in a single month, and one AI startup paid $20,000 for its name on the extension. Defensive registration of your own brand names costs about fifteen dollars each; waiting until a squatter notices you costs whatever they decide to ask. We registered ours the same week.
Koa’s spec sheet is the real news for Salesforce customers
At Dreamforce, Koa was a keynote concept. Now it has numbers: a 120-billion-parameter Mixture-of-Experts model built on NVIDIA’s Nemotron 3 Super, post-trained on a synthetic dataset modeled on nearly thirty years of CRM deployments across more than 14 industries, with a one-million-token context window — and, per Salesforce, no customer data in the training corpus. It is in pilot now, with general availability targeted for winter. Those numbers matter because of how Agentforce pricing works: at roughly ten cents per action under Flex Credits, the economic question is which workflow steps genuinely need a frontier model’s judgment and which can run on a specialized model inside the trust boundary. A CRM-trained model that can hold a million tokens of context — an entire case history, not a summary of one — moves that line. The time to model your routing split is before GA, when you can plan from analysis instead of reacting to an invoice.
Gemini 4: confirmed, not shipped
The other headline circulating this week is Google’s next flagship. The verified status as of October 1: Gemini 4 is in post-training, and Google DeepMind’s leadership says it will ship “much earlier” than the end of the year — with an October 8 “Gemini at Work” event as the speculated venue, though Google has named no date. For Salesforce customers, the practical note is that Gemini is already generally available inside the Agentforce Reasoning Engine, so the model-mix question — which model runs which step — is live today regardless of when Gemini 4 lands. Build the decision framework once, and new models become inputs to it instead of emergencies.
How to read a week like this
- A rename is not a release. Terminology changes alter documents and marketing; capability changes alter architectures. File each story under the right heading before reacting to it.
- Spec sheets beat keynotes. Parameters, context windows, and training-data provenance are decision inputs. Adjectives are not.
- Defensive moves are cheapest while the hype is fresh: a definitions clause in your contracts, your brand names on a surging domain extension, a routing analysis before a model reaches GA.
- Your evals do not care what anything is called. A renamed field and a new flagship model face the same rubric, the same approval gates, and the same acceptance criteria as everything else in production.
This is the lens we bring to engagements: separate the nomenclature from the engineering, quantify what the engineering changes, and leave the naming to the press releases. If the Koa section above describes a decision on your desk, that analysis is what our fixed-scope Agent Economics Review produces. If it is the SI section, your counsel probably owes your contracts a one-line definitions update — and nobody owes anybody a rebrand.